{
  "gradient_accumulation_steps": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_clipping": "auto",
  "zero_allow_untested_optimizer": true,
  "zero_force_ds_cpu_optimizer": false,
  "zero_optimization": {
    "stage": 3,
    "overlap_comm": true,
    "stage3_gather_16bit_weights_on_model_save": true
  },
  "bf16": {
    "enabled": "auto"
  }
}